Weighted-SAMGSR: combining significance analysis of microarray-gene set reduction algorithm with pathway topology-based weights to select relevant genes

Table 7 Performance statistics of selected genes on NSCLC data (multiple-class case)

	Training set (5-fold CV results)				Test set
A. Performance comparison
Method (n)	Error (%)	GBS	BCM	AUPR	Error (%)	GBS	BCM	AUPR
SAMGSR (30)^a	40.7	0.279	0.377	0.462	51.3	0.348	0.407	0.486
W-SAMGSR (27)^a	37.2	0.276	0.378	0.453	51.3	0.345	0.405	0.492
LASSO (95)	38.6	0.281	0.458	0.483	52.7	0.395	0.456	0.485
pSVM (>100)	42.8	0.370	0.344	0.428	53.3	0.433	0.385	0.397
gelnet (>400)	36.6	0.284	0.346	0.416	54.7	0.343	0.377	0.489
RRFE (>200)	36.6	0.272	0.395	0.448	54	0.336	0.410	0.468
B. Performance of the top 3 teams in sbv NSCLC sub-challenge (among 54 teams)
Study (size)	Training data used/Method used				Error (%)	GBS	BCM	AUPR
Ben-Hamo’s (23)	GSE10245, GSE18842, GSE31799/PAM				49.3	--	0.48	0.46
Tarca’s (25)	GSE10245, GSE18842, GSE2109/moderated t-tests + LDA				--	--	0.459	0.454
Tian’s (66)	GSE10245, GSE18842, GSE2109/TGDR in hierarchical way				53.3	0.374	0.440	0.471

Note: W-SAMGSR weighted-SAMGSR, pSVM penalized support vector machine (SCAD penalty term), gelnet generalized elastic net, RRFE reweighted recursive feature elimination, LDA linear discriminant analysis, PAM partitioning around medoid, TGDR threshold gradient descent regularization
^aThe sizes of final model for the stage segmentation because the results for the subtype segmentation for both algorithms are identical (but the final size > 300). Ben-Hamo’s study [31], Tarca’s study [44] and Tian’s study [45] are the 3 best studies in the sbv LC sub-challenge

ISSN: 1745-6150